1. education
2. school

Computers and Student Learning:
Bivariate and Multivariate Evidence on the Availability
and Use of Computers at Home and at School
Thomas Fuchs
Ludger Wößmann
Ifo Working Paper No. 8
May 2005
An electronic version of the paper may be downloaded from the Ifo website: www.ifo.de
Ifo Working Paper No. 8
Computers and Student Learning:
Bivariate and Multivariate Evidence on the Availability
and Use of Computers at Home and at School*
Abstract
We estimate the relationship between computers and students’ educational achievement
in the international student-level PISA database. Bivariate analyses show a positive correlation between achievement and computer availability both at home and at school.
However, once we control extensively for family background and school characteristics,
the relationship gets negative for home computers and insignificant for school computers. Thus, mere availability of computers at home seems to distract students from
effective learning. But achievement shows a positive conditional relationship with computer use for education and communication at home and an inverted U-shaped relationship with computer and internet use at school.
JEL Code: I2.
Keywords: Computers at home, computers at school, student achievement, educational
production, PISA.
Thomas Fuchs
Ifo Institute for Economic Research
at the University of Munich
Poschingerstr. 5
81679 Munich, Germany
Phone: +49(0)89/9224-1317
fuchs@ifo.de
Ludger Wößmann
Ifo Institute for Economic Research
at the University of Munich and CESifo
Poschingerstr. 5
81679 Munich, Germany
Phone: +49(0)89/9224-1699
woessmann@ifo.de
* Financial support by the Volkswagen Foundation for part of this research is gratefully acknowledged.
1. Introduction
Computer use is deemed increasingly important in modern societies. Computers and the
internet have introduced dramatic changes to work processes and to the organization of
corporate structures over the past decade (cf. Lindbeck and Snower 2000). Computers and the
internet have also changed the shopping and recreational behavior of households. Similarly,
students are faced with computers both at home and at school, and governments worldwide
have introduced schemes to equip schools with classroom computers and internet
connections. This paper analyzes whether the availability and use of computers is related to
students’ educational achievement. We use the extensive student-level micro data from the
Programme for International Student Assessment (PISA), an international student
achievement test, to estimate the relationship between computers and student learning
empirically. The PISA database offers information both about the availability and about the
use of computers by students, both at home and at school.
We show that, first, bivariate evidence on the relationship between computers and
students’ educational achievement is highly misleading. Because computer availability at
home is strongly correlated with other family-background characteristics, bivariate results on
computer availability at home are severely biased. Still, simple bivariate correlations are what
many commentators base their assessments on. Even high-quality documents such as the
initial release of the PISA results, albeit cautioning about possible limitations to the
interpretation of bivariate findings, reports the simple bivariate finding that “[s]tudents with
higher values on the index of interest in computers tend to perform better on the combined
reading literacy scale” (OECD 2001, p. 118). We show that the statistically significant
positive correlation between the availability of computers at home and student performance in
math and reading reverses into a statistically significantly negative one as soon as other
family-background influences are extensively controlled for in multivariate regressions.
Second, similar to the case of computer availability at home, bivariate results on computer
availability at school are severely biased because the availability of school computers is
strongly correlated with the availability of other school resources. While the bivariate
correlation between the availability of computers at school and student performance is
strongly and statistically significantly positive, the correlation becomes small and statistically
indistinguishable from zero once other school characteristics are held constant. The
multivariate results illustrate how careless bivariate interpretations can lead to patently false
conclusions.
1
Third, we show that the relationship between computers and student learning differs
strongly between the mere availability of computers and their use as a communicational and
educational device. At home, the negative relationship of student performance with computer
availability contrasts with positive relationships with the use of computers for emailing,
webpage access and the use of educational software. Thus, the mere availability of computers
at home seems to distract students from learning, presumably mainly serving as devices for
playing computer games. Only by using computers in constructive ways can the negative
effect of computer provision on student learning be partly compensated for.
Fourth, the relationship between student achievement and the use of computers and the
internet at school shows an inverted U-shape. That is, students who never use computers or
the internet at school show lower performance than students who sometimes use computers or
the internet at school. But students who use them several times a week perform even lower.
We offer two possible explanations for this pattern. On the one hand, teachers might refrain
from using computers with students of a low ability level. Then, the first part of the pattern
may simply reflect an ability bias, and the second part of the pattern may reflect that computer
use might actually have decreased student learning, as has also been found in a previous
quasi-experimental study (Angrist and Lavy 2002). On the other hand, assuming that there is
no ability bias left after the extensive controls that we include in the regressions, the pattern
might suggest that there is an optimal level of computer and internet use at school,
substantially below a use intensity of several times a week.
Despite the extensive information on family and school background that we can control
for, the PISA study still provides only observational data, where the availability and use of
computers is not randomly divided between a treatment group that has computer access and a
control group that does not have computer access. Therefore, in contrast to randomized
experimental evidence, our evidence has to be interpreted cautiously in terms of descriptive
conditional correlations, which do not necessarily allow for causal inferences because they
may also reflect effects of other, unobservable characteristics. Still, the application shows that
the multivariate analysis can go substantially beyond bivariate correlations in terms of
detecting underlying relationships by disentangling these relationships from other observable
influences. The results illustrate that accounting for other observable influence factors – i.e.,
comparing students who are equal in terms of other observable characteristics – can already
yield results that are perfectly opposite to bivariate patterns.
2
The remainder of the paper is structured as follows. Section 2 briefly describes previous
research on the impact of computers and presents possible hypotheses on the impact of
computerized learning and instruction on students’ educational achievement. Section 3
introduces the database, the identification problem and the empirical model employed.
Section 4 presents the results in terms of computer availability and use at home and at school.
Section 5 concludes.
2. Literature and Hypotheses
2.1
Research on the Economic and Educational Impact of Computers
Using a computer may affect economic outcomes in at least two ways. First, computer skills –
knowing how to use a computer – may have direct effects on productivity and wages. Second,
computers can be used as means for learning other skills, such as math, reading and science,
which in turn may give rise to positive labor-market outcomes.
An increasing literature analyzes the first effect, the direct impact of computer skills on the
labor market. In a seminal paper, Krueger (1993) found that computer use by workers is
related to an expected wage that is approximately 10 to 15 percent higher in cross-sectional
data. However, DiNardo and Pischke (1997) cast doubt that this finding reflects true returns
to computer skills, as similar wage differentials can be found for the use of such devices as
telephones and pencils. Using matched employer-employee panel data to identify the causal
effect of computer skills on wages, Entorf, Gollac and Kramarz (1999) show that the crosssectional wage differential in favor of computer users is nearly entirely due to a selection bias
of high-skilled workers into computer-using employments.
Recently, Borghans and ter Weel (2004) replicate the finding that the ability to effectively
use a computer has no substantial impact on wages. At the same time, they show that math
and writing abilities do yield significant returns on the labor market. Thus, they suggest that
math and writing can be regarded as basic productive skills, while computer skills cannot.
Given that computer skills do not seem to have direct returns on the labor market, but more
basic skills do, the possible impact of computers on learning skills such as math, writing,
reading and science is an interesting topic. The existing evidence focuses on the effect of
classroom computers on student achievement, showing mixed results at best. Reviews of
observational studies such as Cuban (1993), Oppenheimer (1997) and Kirkpatrick and Cuban
(1998) tend towards a negative assessment of the potential of using computers for
instructional purposes in classrooms to improve students’ educational achievement.
3
Wenglinsky (1998) and Angrist and Lavy (2002) even report negative effects of computer
use in schools on some student achievement measures. The quasi-experimental study of
Angrist and Lavy (2002) finds that the introduction of computer-aided instruction in Israeli
schools exerted a statistically significant negative effect on the math achievement of fourthgrade students and a negative but statistically insignificant effect on student achievement in
other subjects and higher grades. Recent studies by Borman and Rachuba (2001) and Rouse
and Krueger (2004) analyze randomized experiments of computerized reading instruction,
finding that noteworthy impacts of computerized instruction can be ruled out.
Thus, the evidence so far does not suggest that computers have a substantial impact on the
economic and educational outcome of individuals, neither in terms of worker wages nor in
terms of student learning. Despite numerous claims by politicians and software vendors to the
contrary, the evidence so far suggests that computer use in schools does not seem to
contribute substantially to students’ learning of basic skills such as math or reading.
2.2
Hypotheses on the Impact of Computers on Student Achievement
At the most basic level, there are both hypotheses suggesting that computers may further
student learning and hypotheses suggesting that computers may hinder student learning.
Therefore, the expected net effect is equivocal – ultimately being an empirical question – and
may depend on factors supporting either the positive or the negative effects.
The positive hypotheses run down to the point that everything else equal, computers
constitute an input in students’ learning process that should help produce better learning
output. Computer use can enhance learning by making education less dependent on differing
teacher quality and by making education available at home throughout the day. Using the
computer to employ educational software can positively infer knowledge to students.
Furthermore, internet access can help students exploit enormous information possibilities for
schooling purposes and increase learning through communication.
A first set of negative hypotheses builds on the idea that, actually, everything else is not
equal. Computerized instruction induces reallocations, substituting alternative, possibly more
effective forms of instruction. Given a constant overall instruction time, this may decrease
student achievement. Also, given that budgets are not perfectly elastic, the introduction of
computerized instruction can result in a reallocation of funds in favor of computers, possibly
substituting more effective instructional materials.
4
A second set of negative hypotheses combines arguments that computers can distract from
learning. This may be particularly salient at home, where computers may be used mainly to
play computer games. This can keep students from doing homework and learning at home.
Survey evidence suggests that computers at home indeed tend to be mainly used as toys
(Wirth and Klieme 2003). Similarly, internet access could offer distraction by chat rooms or
online games, reducing the time spent on doing homework or learning. Thus, the impact on
student learning of the availability of computers and the internet will strongly depend on their
specific uses.
A third set of negative hypotheses surrounds the argument that computer-aided instruction
could restrict the creativity of children. Computerized programs tend to only allow acting in a
predefined way with limited interactive possibilities. This might reduce students’ abilities in
terms of problem solving and creativity, thinking in predetermined schemes but not coming
up with independent creative solutions by their own.
3. Data and Empirical Identification
3.1
The PISA Data
To estimate the relationship between computers and student learning empirically, we use the
student-level dataset of the Programme for International Student Assessment (PISA), an
international student achievement test of 15-year-old students conducted in 2000 by the
Organisation for Economic Co-Operation and Development (OECD). The study tested
student performance in reading, math and science in 32 developed and emerging countries, 28
of which are OECD countries. The OECD ensured a consistent and coherent study design and
as much comparability as possible among the participating countries. The countries
participating in the PISA 2000 study are: Australia, Austria, Belgium, Brazil, Canada, the
Czech Republic, Denmark, Finland, France, Germany, Greece, Hungary, Iceland, Ireland,
Italy, Japan, Korea, Latvia, Liechtenstein,1 Luxembourg, Mexico, the Netherlands, New
Zealand, Norway, Poland, Portugal, the Russian Federation, Spain, Sweden, Switzerland, the
United Kingdom and the United States.2
1
Liechtenstein was not included in our analysis due to lack of internationally comparable country-level
data, e.g. on educational expenditure per student. Note also that there were only 326 15-year-old students in
Liechtenstein in total, 314 of whom participated in PISA.
2
Adams and Wu (2002), OECD (2000, 2001, 2002) and the PISA webpage at www.pisa.oecd.org
provide detailed information on the PISA study.
5
PISA sampled a representative random sample of the population of 15-year-old students in
each country.3 The PISA study tested the students with paper and pencil tests, which lasted
two hours for each student. Using item response theory, the test results were transformed into
test scores with an OECD mean of 500 points and an OECD standard deviation of 100 points.
In this paper, we use the student-level database constructed by Fuchs and Wößmann
(2004), who provide more detailed information and notes on the specific database. They
combine the test results with background information on students and schools from PISA
background questionnaires answered by the specific students and schools tested in PISA. In
addition to the rich PISA data at the student and school level, we also use some country-level
data on the countries’ GDP per capita, on their average educational expenditure per student in
secondary education and on the existence of curriculum-based external exit exams.
In order to use all available information of the PISA test and estimate the maximum
sample of students participating in the test, missing values in the student and school
questionnaires were imputed (see Fuchs and Wößmann 2004 for details on the imputation
method employed). Data imputation is preferable to dropping all observations with a missing
value on at least one variable because it allows a much larger sample size, uses the available
information on other explanatory variables for students with some missing information and
because the estimator using only non-imputed data would yield biased results if the missing
data follow a non-random pattern. To ensure consistency of estimation, the estimated
regressions control for dummies on imputed data (see below).
In this paper, we focus on student performance in math and reading. Math performance is
one of the “basic skills” that Borghans and ter Weel (2004) found to yield productive returns
on the labor market. Similarly, other studies have shown before that math achievement is
most strongly related to productivity (e.g., Bishop 1992). Also, math performance is generally
viewed as being most readily comparable across countries. The sample size of our PISA
database in math is 96,855 students from 31 countries. The PISA 2000 study had a special
focus on the reading literacy of students, where the dataset covers 174,227 students from 31
countries. Thus, we conduct our estimations in math and reading. Science results, for which
we find a pattern very similar to the other two subjects, are not reported here.4
3
Most PISA countries employed a two-stage stratified sampling technique. The first stage drew a
(usually stratified) random sample of schools in which 15-year-old students were enrolled, yielding a minimum
sample of 150 schools per country. The second stage randomly sampled 35 of the 15-year-old students in each
of these schools, with each 15-year-old student in a school having equal probability of selection. Within each
country, this sampling procedure typically led to a sample of between 4,500 and 10,000 tested students.
4
The specific science results are available from the authors on request.
6
Table 1 presents descriptive statistics on the data on computer availability and use at home
and at school that we employ in this paper.5 The table also reports the share of missing and
thus imputed data for each variable. All data on computers at home stem from the student
background questionnaire. In terms of computer availability, students report how many
computers they have at home. We use two dummies on this variable, the first one reporting
whether there is one computer in a student’s home, the second one reporting whether there are
two or more computers in the home. More than half of the students in our sample have more
than one computer in their home, 25% have one computer and 23% do not have a computer at
home. Students also report whether they have internet access at home or not, which 43% of
the students have. Students also report how often they read emails and webpages because they
want to do it, using five answer categories: “never or hardly ever”, “a few times a year”,
“about once a month”, “several times a month” and “several times a week”. We use two
dummies on this in our estimations, one for students who never or hardly ever use emails and
webpages (38% in our sample) and one for students who use them several times a week
(27%), with the remaining students lying in between. A final information on computers at
home is whether students have educational software at home, which 55% of the students
have.
In terms of computer availability at school, school principals report in the school
background questionnaire how much the learning at their school is hindered by not having
enough computers for instruction. Four answer categories are given: “not at all”, “very little”,
“to some extent” and “a lot”. We use two dummies on this variable, one reporting whether
lack of computers hinders learning a lot (12% of the students) and the other whether lack of
computers does not hinder learning at all (30%), with the remaining students again lying in
between. School principals also report the number of computers in their schools, as well as
the number of computers with internet connection. We use both variables in terms of
computers per student. The international mean is 0.126 computers per student at school (or
roughly 8 students on a computer) and 0.067 computers with internet access per student at
school (or roughly 15 students on an internet-connected computer). In terms of computer use
at school, there is additional information in the student background questionnaires. Students
report how often they use computers at their school, in five answer categories: “never or
hardly ever”, “a few times a year”, “about once a month”, “several times a month” and
5
Descriptive statistics on the extensive control variables employed can be found in Fuchs and Wößmann
(2004).
7
“several times a week”. Again, we use the two extreme categories as dummies, where 26% of
the students never or hardly ever use computers at school and 28% use computers at school
several times a week. Similarly, students also report how often they use the internet at their
school, where 49% use it never or hardly ever and only 15% use it several times a week.
3.2
The Identification Problem
As has been shown in numerous studies, students with more advantaged family backgrounds,
such as better-educated parents or parents with higher-paying jobs, generally tend to perform
substantially better in terms of educational achievement (e.g., Wößmann 2004). This may be
the case because of heritability or because parents with more educated backgrounds provide
more inputs for their children’s learning in terms of home instruction, motivation, educational
resources and so on (cf. Behrman and Rosenzweig 2002; Sacerdote 2002; Plug and
Vijverberg 2003). Students from more advantaged family backgrounds also tend to have more
computers at home. Thus, having computers at home will proxy for the economic, social and
educational environment at home at the same time as potentially having an own direct impact
on students’ learning. It follows that any bivariate correlation between computers and student
achievement can well be a sign of other beneficial family-background effects rather than a
sign of computers having an effect on student achievement. This is the essence of the problem
of identifying effects of computer availability and using observational data, because the
computer effects can easily be confounded by effects of other factors. If these other variables
are omitted from the empirical estimation, they will bias the estimated effect of computers.
Therefore, in our estimations, we try to control for as many other observable familybackground effects as possible. In the PISA background questionnaires, we have a vast
amount of information on each student’s personal and family background, as well as on each
school’s resource endowment and organizational structure. By directly controlling for the
economic, social and educational environment at home in a multivariate analysis, we can at
least make sure that any estimated effect of computers will not be driven by these other
observable characteristics.
The identification problem is very similar in the case of computers at schools, because
schools which have more computers also tend to have more of other school resources. For this
reason, we try to substantially control for other school characteristics such as their
endowment with other resources and their institutional features.
8
Still, both in the case of computers at home and computers at school, there will be further
unobserved characteristics left for which we cannot control. If these remaining unobserved
characteristics are correlated with the computer variables and if they are themselves related to
student performance, this will bias our estimated coefficients on the computer variables. For
example, in the case of data on computer use, the decision to use computers may not be
random, but rather endogenously determined by students’ ability. If our control variables do
not fully control for student ability and if this ability is related to measured student
performance, our estimates on computer use may well reflect this ability bias in addition to
any causal effect of computer use. Therefore, our best estimates still do not necessarily show
the causal effect of the computer variables on student performance. Rather, the estimates have
to be interpreted cautiously as descriptive conditional correlations, in the sense that they
report the relationship between computers and student learning conditional on holding
constant the other family-background and school features that we can observe. While this is
substantially more informative than simple bivariate correlations, it is still not the kind of
causal information that controlled experiments could provide.
3.3
The Empirical Model
Following the discussion of the identification problem, we estimate the following multivariate
microeconometric education production function:
Tis = Cis β1 + Bis β 2
(
)
(
)
+ DisC β 3 + DisC Cis β 4 + DisB β 5 + DisB Bis β 6 + ε is
,
(1)
where Tis is the achievement test score of student i in school s. C is the vector of computer
variables, and B is a vector of background data. In the most elaborate specification, this
control vector includes 8 variables on student characteristics, 28 variables on family
background, 12 variables on resource inputs and 12 variables on institutions, as enumerated in
Table 2 (see Fuchs and Wößmann 2004 for details on these control variables). The parameter
vectors β1 to β6 will be estimated in the regression. The inclusion of the imputation dummies
D and the structure of the error term ε will be discussed below. Note that this specification of
the international education production function restricts the effects to be the same in all
countries, as well as at all levels (within schools, between schools and between countries).
While it might be interesting to analyze the potential heterogeneity of certain effects between
countries and between levels, this paper restricts itself to analyzing the international mean
9
effect of computer use on student achievement.6 Fuchs and Wößmann (2004) provide a more
detailed discussion of this general specification of the estimation equation.
We test for the robustness of our results in two further specifications. First, following
Dronkers and Robert (2003), we control for possible school composition effects by including
school means of student gender and of all the 28 family-background variables reported in
Table 2. In effect, this adds 29 additional school-level variables to the background vector B.
By controlling for the socio-economic composition of a school, this specification is meant to
account for possible biases due to the endogenous sorting of students into schools. Second, to
account for potential omitted variables at the country level, we additionally include a whole
set of country dummies in the estimation of equation (1). These country fixed effects control
for unobserved country-specific heterogeneity in the level of student performance.
As discussed in the previous section, some of the data are imputed rather than original.
Generally, data imputation introduces measurement error in the explanatory variables, which
should make it more difficult to observe statistically significant effects. Still, to make sure
that the results are not driven by imputed data, two vectors of dummy variables DC and DB are
included as controls in the estimation. The D vectors contain one dummy for each variable in
the vectors C and B that takes the value of 1 for observations with missing and thus imputed
data and 0 for observations with original data. The inclusion of the D vectors as controls in
the estimation allows the observations with missing data on each variable to have their own
intercepts. Furthermore, the inclusion of the interaction terms between imputation dummies
and data vectors, DCC and DBB, allows them to also have their own slopes for the respective
variables. These imputation controls for each variable with missing values ensure that the
results are robust against possible biases arising from data imputation.
Owing to the complex data structure produced by the PISA survey design and the multilevel nature of the explanatory variables, the error term ε of the regression has a non-trivial
structure. Although we include a considerable amount of school-related variables, we cannot
be sure that there are no omitted variables at the school level. Given the possible dependence
of students within the same school, the use of school-level variables and the fact that schools
6
Wößmann (2003) compares this restricted specification to an alternative two-step specification,
discussing advantages and drawbacks particularly in light of potential omitted country-level variables. He finds
that the substantive results are virtually the same in the alternative specification and provides arguments favoring
the specification employed here. Furthermore, Fuchs and Wößmann (2004) show that this estimation
specification can account for more than 85% of the between-country variation in test scores in each subject.
Therefore, the scope for obvious unobserved country-specific heterogeneity seems small. In the most elaborate
specifications of this paper, we also control for country fixed effects.
10
were the primary sampling unit (PSU) in PISA, there may be unobservable correlation among
the error terms εis at the school level (cf. Moulton 1986 for this problem of hierarchical data
structure). We correct for potential correlations of the error terms by imposing an adequate
structure on the covariance matrix. Thus, we suppose the error term to have the following
structure:
ε is = η s + υ i ,
(2)
where ηs is the school-level element of the error term and υi is the student-specific element of
the error term. We use clustering-robust linear regressions (CRLR) to estimate standard errors
that recognize this clustering of the student-level data within schools. The CRLR method
relaxes the independence assumption and requires only that the observations be independent
across the PSUs, i.e. across schools. To avoid inefficiency due to heteroscedasticity, CRLR
imposes a clustered covariance structure on the covariance matrix, allowing within-school
correlations of the error term. By allowing any given amount of correlation within the PSUs,
CRLR yields consistent and efficient estimates when many observations share the same value
on some but not all independent variables (cf. Deaton 1997; White 1984).
Finally, PISA used a stratified sampling design within each country, producing varying
sampling probabilities for different students. To obtain nationally representative estimates
from the stratified survey data at the within-country level, we employ weighted least squares
(WLS) estimation using sampling probabilities as weights. WLS estimation ensures that the
proportional contribution to the parameter estimates of each stratum in the sample is the same
as would have been obtained in a complete census enumeration (DuMouchel and Duncan
1983; Wooldridge 2001). Furthermore, at the between-country level, our weights give equal
weight to each of the 31 countries.
4. Results
This section reports and discusses the empirical results on the relationship between computers
and student achievement. First, we look at the availability of computers at home and at
school. Then, we look at the use of computers at home and at school.
11
4.1 Computer Availability
Computer Availability at Home
Table 3a reports the results on the availability of computers at home and student performance
in math, successively adding groups of control variables. Column I regresses math
performance only on the two dummies of computer availability at home, without controlling
for any other impact. Similar to the OECD’s (2001, p. 118) finding that a computer interest
index is positively related to performance, computer availability at home is strongly and
statistically significantly positively related to student performance. Students with one
computer at home perform 22.7 achievement points (AP) better than students without a
computer at home, and students with more than one computer at home perform another 6.7
AP better. To provide an impression of the size of this performance difference, it may be
compared to the unconditional international mean difference between students attending 9th
grade and students attending 10th grade. These two grades contain the most students in the
PISA study and show a difference of 30.3 AP in math (33.2 AP in reading). Thus, the
unconditional performance difference between students with several computers at home and
students with no computers at home is approximately equal to a whole grade difference.7
The regression reported in column II adds control variables for student characteristics,
including gender, age and grade, which makes the computer estimates decrease slightly, but
remaining large and statistically significantly positive. In column III, the family-background
controls, such as education and occupation of parents, immigration and family status, are
added to the regression. Now, the coefficients on computer availability at home become small
and statistically insignificantly different from zero. That is, the initial positive correlation
between computer availability at home and student performance was simply driven by the fact
that students from better economic, social and educational family backgrounds tend to have
more computers at home. Holding the other family-background characteristics constant,
computer availability is not related to math performance.
In column IV, control variables for schools’ resource endowments are added to the
regression. Once these are controlled for, computer availability at home is statistically
significantly negatively related to student performance. The more computers there are in a
student’s home, the worse the student’s math performance. This pattern of results gets even
7
As an alternative benchmark, when estimating the average unconditional performance difference per
month between students of different age and extrapolating this to a performance difference per year of age, this
is equal to 12.9 AP in math (16.4 AP in reading).
12
more pronounced when controls for institutional features of the schools are added in column
V. Similarly, it stays unchanged when the schools’ socio-economic composition is controlled
for in column VI, as well as when country fixed effects are included in column VII. In this
most elaborate specification, students with one computer at home perform statistically
significantly 5.9 AP lower than students without a computer at home, and students with
several computers at home perform another 8.0 AP worse.
Table 3b reports the same results for reading literacy. The pattern is the same as in math.
An initial statistically significant and sizable positive correlation between computer
availability at home and reading performance is turned around into a statistically significant
and sizable negative one once student and family background as well as schools’ resources
and institutional features are controlled for.
Computer Availability at School
Tables 4a and 4b report equivalent estimations for the relationship between computer
availability at school and student performance in math and reading. The initial correlations in
column I suggest that students perform statistically significantly and sizably (40.7 AP in math
and 36.5 AP in reading) worse in schools where computers are strongly lacking. There is no
statistically significant performance difference between students in schools without any lack
of computers and students in schools with little or some lack of computers. This pattern of
results stays intact once student and family characteristics are controlled for in columns II and
III. However, the picture starts to turn once measures of schools’ resource endowments are
added as controls. In math, there is no longer a statistically significant performance gap for
students in schools where computers are strongly lacking, and students in schools without a
lack of computers now actually show a statistically significant lower performance. The same
pattern emerges in both subjects in the specifications that additionally control for schools’
institutional features and socio-economic composition in columns V and VI. Once country
fixed effects are added in column VII, none of the performance differences between students
with different extents of computer availability at school is statistically significant. That is, the
initial positive pattern on computer availability at school simply reflects that schools with
better computer availability also feature other positive school characteristics. Once these are
controlled for, computer availability at school is not related to student performance in math
and reading.
Column VIII adds the variable on the number of computers per student at the school to the
most elaborate specification. This variable is also not statistically significantly related to
13
student performance. The statistical insignificance of this variable is already given in a
bivariate correlation, and it remains throughout the other specifications that add more control
variables.
In sum, the results cast strong doubt that the mere availability of computers at home and at
school does a lot to advance students’ educational performance. While bivariate results would
suggest that there is a positive relationship in both cases, these results are spurious. Once
other features of student, family and school background are held constant, computer
availability at home shows a strong statistically significant negative relationship to math and
reading performance, and computer availability at school is unrelated to performance.
The pattern of which particular additional group of control variables brings about the
change in results is telling. In the case of computer availability at home, the first big change
enters when family-background characteristics are controlled for. That is, the strongest bias is
due to the fact that computer availability at home proxies for the economic, social and
educational background of the students’ families. The fact that the negative pattern becomes
more pronounced once schools’ resources, institutions and composition are controlled for
suggests that computer availability at home is also positively correlated with positive school
characteristics. In the case of computer availability at school, the result pattern changes only
once schools’ resource endowments are controlled for. That is, in this case the bias is mainly
due to the fact that computer availability at school proxies for other resource endowments of
the schools, rather than for family background.
4.2 Computer Use
The regressions reported in Tables 5a and 5b extent the results beyond the mere availability
of computers at home or at school. All results are based on the most elaborate specification,
which controls for student characteristics, family background, resource inputs, institutions,
school composition and country fixed effects. The regression in column I simply combines
the availability of computers at home and at school, showing that the previous results are
robust to their joint entrance in the regression. The following regressions add measures of
computer use at home and at school to this specification.
Computer Use at Home
There are three proxy measures for the kind of computer use at home in the PISA
questionnaires, which are added to the regression in column II. The first thing to note is that
14
the negative results on computer availability at home remain unchanged once the computeruse variables are entered.
Only part of the students who have computers at home also have internet access at home.
Having internet access can already hint towards possible alternative uses of computers.
Holding all other influences constant, the performance of students with internet access at
home is statistically significantly better in math and reading than the performance of students
without internet access at home. Additionally, student performance in both math and reading
increases with the frequency of the use of email and webpages by the students. Students who
never or hardly ever read emails and webpages perform statistically significantly worse than
students who use them between a few times a year and several times a month, and students
who use emails and webpages several times a week perform statistically significantly better.
Finally, students that have educational software at home perform statistically significantly
better in math. However, having educational software at home is not statistically significantly
related to student performance in reading literacy.
We suggest two possible interpretations of these positive results of computer use at home.
First, they may simply reflect that more able students tend to be more likely to get internet
access and educational software at home and to use emails and webpages regularly. While
this interpretation is possible, it is not obvious that this kind of ability bias is the main story
here. It might as well be the case that parents tend to be more likely to provide their children
with internet access and educational software if they want to make up for relatively low
ability. Particularly in the case of educational software, parents may tend to buy this
equipment for low-ability rather than high-ability students.
Alternatively, if ability biases do not account for all of the observed performance
differences by computer use, the results may suggest that using computers for productive
purposes at home indeed furthers students’ educational performance. In this interpretation, the
effect of computers at home on student achievement depends on the specific uses to which the
computers are taken. The mere availability of computers at home may in the first instance
serve children as devises to play computer games. This distracts them from learning, and thus
affects their educational performance negatively. But if the computers are instead used for
other means than gaming, namely for communicating by email, accessing information on the
internet and using educational software, this may compensate, at least partly, for the negative
effects induced by computer availability at home and help to advance children’s knowledge in
math and reading.
15
Computer Use at School
Column III of Tables 5a and 5b reports the coefficient estimates on several measures of
computer availability and use at school. Again, adding the computer-use measures does not
change the results on computer availability at school. The mere availability of computers with
internet connection at school is also not statistically significantly related to student
performance.
In terms of computer use at school, on the one hand, students who never or hardly ever use
computers at school perform slightly lower than students who use computers at school
between a few times a year and several times a month. This performance difference is
statistically significant in math, but not in reading. On the other hand, students who use
computers at school several times a week perform sizably and statistically significantly worse
in both math and reading. A very similar pattern of results is found in the case of internet use
at school. In both math and reading, students reporting no internet use at school perform
statistically significantly lower than students of medium internet use at school, while students
who use the internet at school several times a week even perform statistically significantly
worse than both other groups. That is, both in the case of computer use at school and of
internet use at school, the relationship is shaped as an inverted U, with student achievement
initially increasing and subsequently decreasing with the intensity of computer and internet
use at school. The regression in column IV shows that all previous results on computer
availability and use at home and at school are also robust to entering all computer variables at
once.
Again, we suggest two possible patterns of interpretation for the results on computer use at
school. First, there may again be an ability bias. In this case, it seems reasonable to expect
that teachers would not want to use computers with very low-ability students, but only with
students who are reasonably capable to use them. This may explain the fact that students who
never or hardly ever use computers or the internet at school perform somewhat lower, which
then would simply reflect that they are low-ability students. It seems less likely that such an
ability effect can also account for the substantially lower performance of students who use
computers and the internet at school several times a week. This latter finding might instead be
explained by a true negative effect of excessive computer use at school. As argued above,
computerized instruction may substitute alternative, more effective forms of instruction, and it
may also harm the creativity of children’s learning. This result would then be in line with
16
Angrist and Lavy’s (2002) quasi-experimental finding that in some cases, computer use at
schools can have negative effects on students’ educational achievement.
There is also a second interpretation of the pattern of results, which does not consider
ability biases. In this interpretation, the inverted U-shape of the relationship between
computer and internet use at school and student performance may be due to a true causal
effect of computer and internet use. In this case, some computerized instruction would be
beneficial for student learning, constituting a valuable input in the students’ learning process.
Only at higher intensities of computer and internet use at school would the negative effects of
computer and internet use set in, in terms of crowding out more effective methods of teaching
and of hindering student creativity. Thus, there may be an optimal level of computer and
internet use at school, with student learning initially increasing and subsequently decreasing
with the intensity of computer and internet use. The presented results suggest that this optimal
level may be pretty low, though, somewhere between using computers and the internet at
school “a few times a year” and “several times a month”, as students in these categories
perform better than students in the categories of both “never or hardly ever” and “several
times a week”.
5. Conclusions
This paper has found that despite bivariate correlations that show a positive relationship, once
family background and school characteristics are extensively controlled for, the mere
availability of computers at home is negatively related to student performance in math and
reading, and the availability of computers at school is unrelated to student performance. By
contrast, student performance is positively related to the use of computers at home for
accessing emails and webpages and to the availability of educational software at home.
Finally, student performance shows an inverted U-shaped relationship with the extent of
computer and internet use at school, rising with some use but falling again with a use of
several times a week.
Despite the extensive use of control variables, the analysis has still been descriptive rather
than causal. For a thorough analysis of causal effects of computers on student performance,
we have to be sure that the variation in computer availability and use on which the analysis is
based is truly exogenous to the model. For analyses based on observational rather than
experimental data, this is always a somewhat strong assumption. However, given our
extensive controlling for student, family and school background effects, the descriptive results
17
that we obtain give a clearer picture of the relationship between computers and student
performance than mere bivariate analyses. Actually, the diametrically opposite difference
between our bivariate and multivariate results for the availability of computers at home shows
that the simple bivariate picture is clearly not based on exogenous variation in computer
availability, and is therefore a far cry from depicting any causal effect of computer
availability. Rather than the statistically significant and large positive bivariate correlation
between computer availability and student performance, once we control extensively for
family background we find a statistically significant and large negative relationship between
the two. This is still a descriptive finding, but one which we think should come much closer to
any causal effect than the often presented bivariate relationship.
Our results on computer availability and use at school corroborate previous work on school
computers such as Angrist and Lavy (2002) and Rouse and Krueger (2004), who also find
disappointing results in terms of effects on students’ educational performance. Our results on
computer availability and use at home extends this evidence, illustrating that there is also a
negative relationship between home computer availability and student achievement, but a
positive relationship between home computer use for internet communication and educational
software. Similarly, our results on internet use at school complement the previous evidence on
computer use at school.
Having a computer at home and using it at school will almost certainly raise some
computer skills. What our results suggest is only that this may come at the expense of other
skills. However, the results in Borghans and ter Weel (2004) show that these other (math and
writing) skills are the ones that yield significant labor-market returns, not the computer skills.
Our results also cast strong doubt on the possibility of giving a causal interpretation to
bivariate results for other variables. For example, the OECD (2001) reports bivariate
correlations of student performance with such features as reading interest, motivation,
engagement and different teaching techniques. Our results suggest that any such finding may
well be spurious, being driven by other important factors. In this sense, our exercise provides
an illustration of the need for, at least, multivariate analysis.
18
References
Adams, Ray, Margaret Wu (eds.) (2002). PISA 2000 Technical Report. Paris: Organisation
for Economic Co-operation and Development (OECD).
Angrist, Joshua, Victor Lavy (2002). New Evidence on Classroom Computers and Pupil
Learning. Economic Journal 112 (482): 735-765.
Behrman, Jere R., Mark R. Rosenzweig (2002). Does Increasing Women’s Schooling Raise
the Schooling of the Next Generation? American Economic Review 92 (1): 323-334.
Bishop, John H. (1992). The Impact of Academic Competencies on Wages, Unemployment,
and Job Performance. Carnegie-Rochester Conference Series on Public Policy 37: 127194.
Borghans, Lex, Bas ter Weel (2004). Are Computer Skills the New Basic Skills? The Returns
to Computer, Writing and Math Skills in Britain. Labour Economics 11 (1): 85-98.
Borman, Geoffrey D., Laura T. Rachuba (2001). Evaluation of the Scientific Learning
Corporation’s FastForWord Computer-Based Training Program in the Baltimore City
Public Schools. Report Prepared for the Abell Foundation.
Cuban, Larry (1993). Computers Meet Classroom: Classroom Wins. Teachers College
Record 95 (2): 185-210.
Deaton, Angus (1997). The Analysis of Household Surveys: A Microeconometric Approach to
Development Policy. Baltimore: The Johns Hopkins University Press.
DiNardo, John E., Jörn-Steffen Pischke (1997). The Returns to Computer Use Revisited:
Have Pencils Changed the Wage Structure Too? Quarterly Journal of Economics 112 (1):
291-303.
Dronkers, Jaap, Péter Robert (2003). The Effectiveness of Public and Private Schools from a
Comparative Perspective. EUI Working Paper SPS 2003-13. Florence: European
University Institute.
DuMouchel, William H., Greg J. Duncan (1983). Using Sample Survey Weights in Multiple
Regression Analyses of Stratified Samples. Journal of the American Statistical Association
78 (383): 535-543.
Entorf, Horst, Michel Gollac, Francis Kramarz (1999). New Technologies, Wages, and
Worker Selection. Journal of Labor Economics 17 (3): 464-491.
Fuchs, Thomas, Ludger Wößmann (2004). What Accounts for International Differences in
Student Performance? A Re-examination using PISA Data. CESifo Working Paper (in
press). Munich: CESifo.
Kirkpatrick, Heather, Larry Cuban (1998). Computers Make Kids Smarter – Right? Technos
Quarterly 7 (2). [http://www.technos.net/tq_07/2cuban.htm]
Krueger, Alan B. (1993). How Computers Have Changed the Wage Structure: Evidence from
Microdata, 1984-1989. Quarterly Journal of Economics 108 (1): 33-60.
Lindbeck, Assar, Dennis Snower (2000). Multitask Learning and the Reorganization of Work:
From Tayloristic to Holistic Organization. Journal of Labor Economics 18 (3): 353-376.
Moulton, Brent R. (1986). Random Group Effects and the Precision of Regression Estimates.
Journal of Econometrics 32 (3): 385-397.
Oppenheimer, Todd (1997). The Computer Delusion. Atlantic Monthly 280 (1), July.
19
Organisation for Economic Co-operation and Development (OECD) (2000). Measuring
Student Knowledge and Skills: The PISA 2000 Assessment of Reading, Mathematical and
Scientific Literacy. Paris: OECD.
Organisation for Economic Co-operation and Development (OECD) (2001). Knowledge and
Skills for Life: First Results from the OECD Programme for International Student
Assessment (PISA) 2000. Paris: OECD.
Organisation for Economic Co-operation and Development (OECD) (2002). Manual for the
PISA 2000 Database. Paris: OECD.
Plug, Erik, Wim Vijverberg (2003). Schooling, Family Background, and Adoption: Is It
Nature or Is It Nurture? Journal of Political Economy 111 (3): 611-641.
Rouse, Cecilia E., Alan B. Krueger, with Lisa Markman (2004). Putting Computerized
Instruction to the Test: A Randomized Evaluation of a “Scientifically-based” Reading
Program. NBER Working Paper 10315. Cambridge, MA: National Bureau of Economic
Research.
Sacerdote, Bruce (2002). The Nature and Nurture of Economic Outcomes. American
Economic Review, Papers and Proceedings 92 (2): 344-348.
Wenglinsky, Harold (1998). Does It Compute? The Relationship Between Educational
Technology and Achievement in Mathematics. Princeton, NJ: Policy Information Center,
Research Division, Educational Testing Service.
White, Halbert (1984). Asymptotic Theory for Econometricians. Orlando: Academic Press.
Wirth, Joachim, Eckhard Klieme (2003). Computernutzung. In: Deutsches PISA-Konsortium
(ed.), PISA 2000: Ein differenzierter Blick auf die Länder der Bundesrepublik
Deutschland: 195-209. Opladen: Leske + Budrich.
Wooldridge, Jeffrey M. (2001). Asymptotic Properties of Weighted M-Estimators for
Standard Stratified Samples. Econometric Theory 17 (2): 451-470.
Wößmann, Ludger (2003). Schooling Resources, Educational Institutions, and Student
Performance: The International Evidence. Oxford Bulletin of Economics and Statistics 65
(2): 117-170.
Wößmann, Ludger (2004). How Equal Are Educational Opportunities? Family Background
and Student Achievement in Europe and the United States. CESifo Working Paper 1162.
Munich: CESifo.
20
Table 1: Descriptive Statistics on Achievement and Computers
Mean
Test scores
Math
Reading
Computers at home
Computers at home
None
One
More than one
Internet access at home
496.1
495.4
Std. Dev.
102.6
101.3
Source
Imputed
St
St
0.000
0.000
0.229
0.247
0.524
0.431
St
St
St
St
0.025
0.025
0.025
0.024
0.381
0.265
0.552
St
St
St
0.036
0.036
0.028
0.115
0.304
0.126
0.385
Sc
Sc
Sc
0.044
0.044
0.116
0.218
St
St
Sc
0.036
0.036
0.137
St
St
0.039
0.039
Use of email and webpages
Never or hardly ever
Several times a week
Educational software at home
Computers at school
Computers at school
Strongly lacking
No lack at all
PCs per student
Computer use at school
Never or hardly ever
Several times a week
PCs with internet access per student
0.259
0.277
0.067
Internet use at school
Never or hardly ever
Several times a week
0.485
0.152
Notes: Mean: International mean, based on non-imputed data for each variable,
weighted by sampling probabilities. – Std. Dev.: International standard deviation
(only for discrete variables). – Source: Data source and thus level of observation: St =
student achievement test or student background questionnaire; Sc = school
background questionnaire. – Imputed: Fraction of students with missing and thus
imputed data, weighted by sampling probabilities.
21
Table 2: Control Variables
Student characteristics [8]
Family background [28]
• Gender
• Parental education (5 dummies)
• Age
• Migration status of father, mother and
student (3 dummies)
• Grade (6 dummies)
• Family status (3 dummies)
• Parents’ work status (3 dummies)
• Parental occupation (2 dummies)
• Number of books at home (6 dummies)
• School’s community location
(5 dummies)
• GDP per capita of country
Resource inputs [12]
Institutions [12]
• Class size in subject (instrumented
by school’s student-teacher ratio)
• External exit exams
• Educational expenditure per student
of country
• School autonomy in determining course
content, choosing textbooks, formulating
school budget, deciding on budget allocations,
hiring teachers, firing teachers, establishing
teachers’ starting salaries and establishing
teachers’ salary increases (8 dummies)
• Instructional material (2 dummies)
• Teacher education (3 dummies)
• Instruction time
• Homework time in subject
(2 dummies)
• Parental support (2 dummies)
22
• Standardized tests
• Public vs. private school management
• Share of government funding in school budget
Table 3a: Computer Availability at Home and Student Performance in Math
I
Computers at home
One
More than one
Student characteristics [8]
Family background [28]
Resource inputs [12]
Institutions [12]
School composition [29]
Country dummies [30]
R2
Observations
22.737*** (1.589)
29.452*** (1.604)
–
–
–
–
–
–
0.03
96,855
II
17.143*** (1.457)
21.693*** (1.492)
incl.
–
–
–
–
–
0.14
96,855
III
1.791 (1.290)
-2.068 (1.355)
incl.
incl.
–
–
–
–
0.26
96,855
IV
V
-2.103* (1.216)
-6.576*** (1.266)
incl.
incl.
incl.
–
–
–
0.31
96,855
-3.921*** (1.174)
-9.804*** (1.192)
incl.
incl.
incl.
incl.
–
–
0.33
96,855
VI
-5.580*** (1.121)
-12.118*** (1.125)
incl.
incl.
incl.
incl.
incl.
–
0.37
96,855
VII
-5.907*** (1.106)
-13.863*** (1.070)
incl.
incl.
incl.
incl.
incl.
incl.
0.39
96,855
Notes: Dependent variable: PISA international math test score. – I-II: WLS regressions. – III-VII: 2SLS regressions with class size instrumented by schools’
student-teacher ratio. – Regressions weighted by students’ sampling probabilities. – Clustering-robust standard errors (taking account of correlated error terms
within schools) in parentheses. – Significance level (based on clustering-robust standard errors): *** 1 percent. – * 10 percent.
23
Table 3b: Computer Availability at Home and Student Performance in Reading
I
Computers at home
One
More than one
Student characteristics [8]
Family background [28]
Resource inputs [12]
Institutions [12]
School composition [29]
Country dummies [30]
R2
Observations
22.011*** (1.298)
26.008*** (1.427)
–
–
–
–
–
–
0.02
174,227
II
17.047*** (1.165)
19.205*** (1.285)
incl.
–
–
–
–
–
0.15
174,227
III
IV
3.783*** (1.016)
-1.786 (1.184)
incl.
incl.
–
–
–
–
0.29
174,227
-1.075 (1.126)
-5.885*** (1.246)
incl.
incl.
incl.
–
–
–
0.33
174,227
V
VI
-4.227*** (1.026) -5.558*** (0.893)
-10.381*** (1.055) -12.500*** (0.914)
incl.
incl.
incl.
incl.
incl.
incl.
incl.
incl.
–
incl.
–
–
0.34
0.38
174,227
174,227
VII
-6.551*** (0.817)
-15.606*** (0.819)
incl.
incl.
incl.
incl.
incl.
incl.
0.40
174,227
Notes: Dependent variable: PISA international reading test score. – I-II: WLS regressions. – III-VII: 2SLS regressions with class size instrumented by schools’
student-teacher ratio. – Regressions weighted by students’ sampling probabilities. – Clustering-robust standard errors (taking account of correlated error terms
within schools) in parentheses. – Significance level (based on clustering-robust standard errors): *** 1 percent.
24
Table 4a: Computer Availability at School and Student Performance in Math
I
Computers at school
Strongly lacking
-40.717*** (3.600)
No lack at all
-0.280 (2.403)
PCs per student
–
Student characteristics [8]
–
Family background [28]
–
Resource inputs [12]
–
Institutions [12]
–
School composition [29]
–
Country dummies [30]
–
0.02
R2
Observations
96,855
II
III
IV
-31.847*** (3.170)
1.420 (2.081)
–
incl.
–
–
–
–
–
0.13
96,855
-15.611*** (2.466)
1.659 (1.805)
–
incl.
incl.
–
–
–
–
0.26
96,855
V
-2.777 (2.307)
-1.568 (2.133)
**
(1.733)
-4.395
-6.053*** (1.640)
–
–
incl.
incl.
incl.
incl.
incl.
incl.
–
incl.
–
–
–
–
0.31
0.33
96,855
96,855
VI
VII
-0.463 (2.015) -1.592 (1.870)
-2.344* (1.415) -1.314 (1.236)
–
–
incl.
incl.
incl.
incl.
incl.
incl.
incl.
incl.
incl.
incl.
–
incl.
0.36
0.39
96,855
96,855
VIII
-1.624 (1.872)
-1.315 (1.236)
-0.480 (0.829)
incl.
incl.
incl.
incl.
incl.
incl.
0.39
96,855
Notes: Dependent variable: PISA international math test score. – I-II: WLS regressions. – III-VIII: 2SLS regressions with class size instrumented by schools’ student-teacher
ratio. – Regressions weighted by students’ sampling probabilities. – Clustering-robust standard errors (taking account of correlated error terms within schools) in parentheses. –
Significance level (based on clustering-robust standard errors): *** 1 percent. – ** 5 percent. – * 10 percent.
25
Table 4b: Computer Availability at School and Student Performance in Reading
I
Computers at school
Strongly lacking
-36.456*** (3.285)
No lack at all
-0.092 (2.310)
PCs per student
–
Student characteristics [8]
–
Family background [28]
–
Resource inputs [12]
–
Institutions [12]
–
School composition [29]
–
Country dummies [30]
–
0.01
R2
Observations
174,227
II
III
IV
-27.819*** (2.834)
1.010 (1.953)
–
incl.
–
–
–
–
–
0.15
174,227
-14.862*** (2.155)
0.407 (1.592)
–
incl.
incl.
–
–
–
–
0.29
174,227
V
-5.213** (2.285)
-2.868 (2.069)
***
(1.693)
-5.076
-6.234*** (1.564)
–
–
incl.
incl.
incl.
incl.
incl.
incl.
–
incl.
–
–
–
–
0.32
0.34
174,227
174,227
VI
VII
-1.845 (1.806) -1.778 (1.631)
-2.430* (1.306) -1.184 (1.108)
–
–
incl.
incl.
incl.
incl.
incl.
incl.
incl.
incl.
incl.
incl.
–
incl.
0.38
0.40
174,227
174,227
VIII
-1.812 (1.634)
-1.179 (1.108)
-0.584 (0.744)
incl.
incl.
incl.
incl.
incl.
incl.
0.40
174,227
Notes: Dependent variable: PISA international reading test score. – I-II: WLS regressions. – III-VIII: 2SLS regressions with class size instrumented by schools’ student-teacher
ratio. – Regressions weighted by students’ sampling probabilities. – Clustering-robust standard errors (taking account of correlated error terms within schools) in parentheses. –
Significance level (based on clustering-robust standard errors): *** 1 percent. – ** 5 percent. – * 10 percent.
26
Table 5a: Computer Use at Home and at School and Student Performance in Math
I
Computers at home
One
-5.895*** (1.106)
More than one
-13.843*** (1.069)
Internet access
–
Use of email and webpages
Never or hardly ever
–
Several times a week
–
Educational software
–
Computers at school
Strongly lacking
-1.803 (1.864)
No lack at all
-1.218 (1.233)
PCs per student
–
Computer use at school
Never or hardly ever
–
Several times a week
–
PCs with internet access per student
–
Internet use at school
Never or hardly ever
–
Several times a week
–
Student characteristics [8]
incl.
Family background [28]
incl.
Resource inputs [12]
incl.
Institutions [12]
incl.
School composition [29]
incl.
Country dummies [30]
incl.
0.40
R2
96,855
Observations
II
III
IV
-6.981*** (1.103)
-16.359*** (1.080)
4.859*** (0.955)
–
–
–
-7.086*** (1.104)
-16.164*** (1.078)
4.590*** (0.935)
-7.614*** (0.864)
4.055*** (0.884)
2.116*** (0.815)
–
–
–
-7.003*** (0.868)
6.301*** (0.914)
1.788** (0.806)
–
–
–
-0.871 (1.887)
-1.357 (1.221)
1.415 (2.073)
-1.093 (1.882)
-1.259 (1.209)
1.575 (2.066)
–
–
–
-1.938* (1.039)
-6.321*** (1.098)
-4.134 (3.874)
-2.323** (1.035)
-6.075*** (1.089)
-4.436 (3.914)
–
–
incl.
incl.
incl.
incl.
incl.
incl.
0.40
96,855
-3.741*** (1.024)
-6.694*** (1.294)
incl.
incl.
incl.
incl.
incl.
incl.
0.40
96,855
-2.079** (1.023)
-9.224*** (1.336)
incl.
incl.
incl.
incl.
incl.
incl.
0.40
96,855
Notes: Dependent variable: PISA international math test score. – 2SLS regressions with class size instrumented by
schools’ student-teacher ratio. – Regressions weighted by students’ sampling probabilities. – Clustering-robust
standard errors (taking account of correlated error terms within schools) in parentheses. – Significance level
(based on clustering-robust standard errors): *** 1 percent. – ** 5 percent. – * 10 percent.
27
Table 5b: Computer Use at Home and at School and Student Performance in Reading
I
Computers at home
One
-6.603*** (0.815)
More than one
-15.697*** (0.820)
Internet access
–
Use of email and webpages
Never or hardly ever
–
Several times a week
–
Educational software
–
Computers at school
Strongly lacking
-1.960 (1.622)
No lack at all
-1.149 (1.102)
PCs per student
–
Computer use at school
Never or hardly ever
–
Several times a week
–
PCs with internet access per student
–
Internet use at school
Never or hardly ever
–
Several times a week
–
Student characteristics [8]
incl.
Family background [28]
incl.
Resource inputs [12]
incl.
Institutions [12]
incl.
School composition [29]
incl.
Country dummies [30]
incl.
0.40
R2
Observations
174,227
II
III
IV
-7.547*** (0.815)
-17.995*** (0.836)
4.808*** (0.705)
–
–
–
-7.631*** (0.811)
-17.932*** (0.834)
4.450*** (0.697)
-6.403*** (0.665)
5.916*** (0.668)
0.088 (0.603)
–
–
–
-6.183*** (0.665)
8.703*** (0.681)
-0.278 (0.596)
–
–
–
-1.166 (1.642)
-1.216 (1.092)
-1.798 (2.537)
-1.363 (1.626)
-1.152 (1.075)
-1.540 (2.398)
–
–
–
-0.528 (0.812)
-5.067*** (0.851)
2.202 (4.790)
-0.887 (0.806)
-4.934*** (0.843)
1.688 (4.648)
–
–
incl.
incl.
incl.
incl.
incl.
incl.
0.40
174,227
-2.986*** (0.809)
-9.552*** (1.021)
incl.
incl.
incl.
incl.
incl.
incl.
0.40
174,227
-1.519* (0.798)
-12.577*** (1.037)
incl.
incl.
incl.
incl.
incl.
incl.
0.41
174,227
Notes: Dependent variable: PISA international reading test score. – 2SLS regressions with class size instrumented
by schools’ student-teacher ratio. – Regressions weighted by students’ sampling probabilities. Clustering-robust
standard errors (taking account of correlated error terms within schools) in parentheses. – Significance level
(based on clustering-robust standard errors): *** 1 percent. – * 10 percent.
28
Ifo Working Papers
No. 7
Werding, M., Survivor Benefits and the Gender Tax-Gap in Public Pension Schemes –
Work Incentives and Options for Reform, May 2005.
No. 6
Holzner, Chr., Search Frictions, Credit Constraints and Firm Financed General Training,
May 2005.
No. 5
Sülzle, K., Duopolistic Competition between Independent and Collaborative Business-toBusiness Marketplaces, March 2005.
No. 4
Becker, Sascha O., K. Ekholm, R. Jäckle and M.-A. Muendler, Location Choice and Employment Decisions: A Comparison of German and Swedish Multinationals, March 2005.
No. 3
Bandholz, H., New Composite Leading Indicators for Hungary and Poland, March 2005.
No. 2
Eggert, W. and M. Kolmar, Contests with Size Effects, January 2005.
No. 1
Hanushek, E. and L. Wößmann, Does Educational Tracking Affect Performance and Inequality? Differences-in-Differences Evidence across Countries, January 2005.